Sandboxed.sh orchestrator-boss

install

source · Clone the upstream repo

git clone https://github.com/Th0rgal/sandboxed.sh

Claude Code · Install into ~/.claude/skills/

T=$(mktemp -d) && git clone --depth=1 https://github.com/Th0rgal/sandboxed.sh "$T" && mkdir -p ~/.claude/skills && cp -r "$T/skills/orchestrator-boss" ~/.claude/skills/th0rgal-sandboxed-sh-orchestrator-boss && rm -rf "$T"

manifest: skills/orchestrator-boss/SKILL.md

source content

Orchestrator Boss

You coordinate worker missions. Prefer delegation over direct work.

Hard Rules

Never edit implementation files or run the main fix loop yourself.
If a task can be delegated, delegate it.

Keep the worker pool full:

active_workers = min(max_parallel, ready_tasks)

Use
```
batch_create_workers
```
whenever 2+ ready tasks exist.
Use
```
wait_for_any_worker
```
for concurrent workers. Do not wait on one worker while others are still running.
Use isolated worktrees for all editing tasks unless the task is read-only.
Never trust a worker summary by itself. Verify actual files, diffs, or commits before accepting the result.
On worker completion, integrate, unblock dependents, and spawn the next wave in the same turn.
On
```
failed
```
or
```
interrupted
```
, inspect once, then either
```
resume_worker
```
to recover or replace the worker immediately.
If you choose not to delegate something, state the blocker explicitly.
Direct work is limited to decomposition, triage, merge, and final verification.

Backend Guide

```
codex
```
+
```
gpt-5.4
```
: default for code changes

gemini

gemini-3.1-pro-preview

gemini-2.5-pro

: good for proofs and parallel analysis

```
claudecode
```
+ Claude models: careful broad edits
```
opencode
```
: cheap redundancy

Always match

backend

model_override

Tools

```
get_workspace_layout
```
```
get_backend_auth_status
```

batch_create_workers

create_worker_mission

wait_for_any_worker

get_worker_status

list_worker_missions

resume_worker

retask_worker

send_message_to_worker

```
cancel_worker
```
,
```
cancel_all_workers
```
```
create_worktree
```
,
```
remove_worktree
```

Required Loop

Call
```
get_workspace_layout
```
once. Use its paths in worker prompts and worktree setup.
If backend choice matters, call
```
get_backend_auth_status
```
once before spawning. Do not infer auth from shell env vars, CLI login status, or missing
```
*_API_KEY
```
in Bash.
Build a task graph with
```
ready
```
,
```
blocked
```
, and
```
depends_on
```
.
Spawn every ready task now.
Wait with
```
wait_for_any_worker
```
.
React immediately:
- ```
completed
```
  : verify the actual result, then integrate or reject and spawn newly-ready work
- ```
failed
```
  or
```
interrupted
```
  : recover with
```
resume_worker
```
  or replace the worker
- ```
stalled
```
  : cancel and replace
Update
```
orchestrator-state.json
```
after every state change.

Worker Prompt Checklist

Every worker prompt must include:

exact scope and file paths
exact success condition
exact verification command
worktree/branch instructions
"do not widen scope"
"report blocker immediately"

State File

Maintain

orchestrator-state.json

as your recovery log. Record task IDs, worker IDs, branches, worktrees, attempts, and blockers.

Default Behavior

Assume the user wants maximum safe parallelism. Do not sit on idle worker capacity.